app.py

import ollama
import gradio

history = []


def get_history_messages():
    messages = []
    for user, assist in history:
        messages.append({"role": "user", "content": user})
        messages.append({"role": "assistant", "content": assist})
    return messages


def predict(prompt):
    response = ollama.chat(
        model="tinydolphin",
        messages=[
            *get_history_messages(),
            {"role": "user", "content": prompt}
        ],
        stream=True
    )
    history.append((prompt, ""))
    message = ""
    for chunk in response:
        message += chunk["message"]["content"]
        history[-1] = (prompt, message)
        yield "", history

with gradio.Blocks(fill_height=True) as demo:
    chat = gradio.Chatbot(scale=1)
    with gradio.Row(variant="compact"):
        prompt = gradio.Textbox(show_label=False, scale=6, autofocus=True)
        button = gradio.Button(scale=1)

    for handler in [button.click, prompt.submit]:
        handler(predict, inputs=[prompt], outputs=[prompt, chat])


if __name__ == '__main__':
    demo.launch()